import tqdm
import cv2
import numpy as np
from deep_privacy.inference.anonymizer import Anonymizer
from deep_privacy.inference import utils as inference_utils
class SimpleAnonymizer(Anonymizer):
def __init__(self, *args, **kwargs):
super().__init__(kwargs)
def anonymize_face(self, face):
raise NotImplementedError
def anonymize_images(self, images, im_bboxes, im_keypoints=None, max_face_size=1.0):
anonymized_images = []
for im_idx, im in enumerate(tqdm.tqdm(images, desc="Anonymizing images")):
anonymized_image = im.copy()
bboxes = im_bboxes[im_idx]
bboxes = inference_utils.filter_bboxes(bboxes,
anonymized_image.shape,
max_face_size)
for bbox in bboxes:
x0, y0, x1, y1 = [int(_) for _ in bbox]
x0, y0 = max(0, x0), max(0, y0)
x1, y1 = min(im.shape[1], x1), min(im.shape[0], y1)
if y1 - y0 <= 0 or x1 - x0 <= 0:
continue
face = im[y0:y1, x0:x1]
if face.shape[0] == 0 or face.shape[1] == 0:
continue
orig_shape = face.shape
face = self.anonymize_face(face)
assert orig_shape == face.shape, f"Did not return equal sized face."
anonymized_image[y0:y1, x0:x1] = face
anonymized_images.append(anonymized_image)
return anonymized_images
class PixelationAnonymizer(SimpleAnonymizer):
def __init__(self, pixelation_size, **kwargs):
super().__init__(kwargs)
self.pixelation_size = pixelation_size
print("Pixelation initialize!")
def anonymize_face(self, face):
orig_size = face.shape
face = cv2.resize(face, (self.pixelation_size,
self.pixelation_size))
face = cv2.resize(face, (orig_size[1], orig_size[0]))
return face
class BlurAnonymizer(SimpleAnonymizer):
def __init__(self, blur_type, **kwargs):
super().__init__(kwargs)
self.blur_type = blur_type
print("Blur anonymizer initialize!")
def anonymize_face(self, face):
if self.blur_type == "heavy_blur":
ksize = int(0.3 * face.shape[1])
ksize = max(1, ksize)
face = cv2.blur(face, (ksize, ksize))
return face
elif self.blur_type == "gaussian_blur":
face = cv2.GaussianBlur(face, (9, 9), sigmaX=3, sigmaY=3)
return face
else:
raise AttributeError(f'Undefined blur type: {self.blur_type}')
class BlackOutAnonymizer(SimpleAnonymizer):
def __init__(self, *args, **kwargs):
super().__init__(kwargs)
print("BlackOut anonymizer initialize!")
def anonymize_face(self, face):
return np.random.normal(
loc=face.mean(), scale=face.std(),
size=face.shape
)
